Abstract: This paper examined the basic structure and training method of Faster R-CNN (convolutional neural networks) target detection network. The state data set of mechanical and electrical equipment was established, and the target detection network was trained. In a single step, the region of pointer instrument could be extracted, and the reading of digital instrument and the state of switch and plug could be recognized. The target detection network was tested under different viewing angles and illumination intensities. The results show that the model can maintain the accuracy of more than 90% in different environments. Finally, based on the reasoning results, the intelligent maintenance assistant system for mechanical and electrical equipment developed based on Unity 3D software and HoloLens 2 hardware was applied to the retrieval of the mixed reality (MR) holographic induction maintenance information, thus guiding the operation of the support personnel. In order to verify the availability of the system, the experimental verification process was added, and the experimental results show that the experimenter could complete the maintenance task quickly and efficiently using MR. In addition, test and evaluation were conducted based on the operation time and questionnaire survey, and qualitative analysis was carried out regarding the advantages of the system. 

Key words: mixed reality, faster R-CNN, intelligent fault diagnosis, maintenance assistance, experimental verification